Um estudo comparativo entre redes neurais convolucionais para a classificação de imagens

Abstract

The recognition of objects in images is becoming easier for the computers, due to the advances in studies of computational vision, that uses neural networks of deep learning. These networks had big advances since 2012 and have been used on different problems of classifications that evolve images. In this work, are made three implementations of convolutional neural networks, that are: VGG-16, RestNet-50 and Inception. The purpose of this work is making a comparative of the implementations, evaluating which one is better to solve a problem of image classification, considering some metrics, for the competition of classification of the Simpsons’ characters. Between the chosen metrics, the better accuracy result was the ResNet50 with 93.4% and the best result of log loss was the Inception with 0.26, that showed that the networks were able to solve the problem just fine.